GPU Acceleration in LAMMPS

Minimizing the Code Burden • Advantages • Allow a legacy code to take advantage of accelerators without rewri1ng the en1re thing • In some cases, there is no advantage to por1ng • Mul1-­‐Core CPUs can be compe11ve with GPU accelerators for some rou1nes (latency-­‐bound, thread divergence, etc.), especially at lower par1cle counts • If concurrent CPU/GPU calcula1ons are used effec1vely, there can be no advantage to por1ng certain rou1nes • For some simula1ons, the upper bound for performance improvements due to por1ng addi1onal rou1nes and removing data transfer is < 5%.